Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2013 Davidlohr Bueso <davidlohr@hp.com>
4 *
5 * futex-hash: Stress the hell out of the Linux kernel futex uaddr hashing.
6 *
7 * This program is particularly useful for measuring the kernel's futex hash
8 * table/function implementation. In order for it to make sense, use with as
9 * many threads and futexes as possible.
10 */
11
12/* For the CLR_() macros */
13#include <string.h>
14#include <pthread.h>
15
16#include <errno.h>
17#include <signal.h>
18#include <stdlib.h>
19#include <linux/compiler.h>
20#include <linux/kernel.h>
21#include <linux/zalloc.h>
22#include <sys/time.h>
23#include <internal/cpumap.h>
24#include <perf/cpumap.h>
25
26#include "../util/stat.h"
27#include <subcmd/parse-options.h>
28#include "bench.h"
29#include "futex.h"
30
31#include <err.h>
32
33static unsigned int nthreads = 0;
34static unsigned int nsecs = 10;
35/* amount of futexes per thread */
36static unsigned int nfutexes = 1024;
37static bool fshared = false, done = false, silent = false;
38static int futex_flag = 0;
39
40struct timeval start, end, runtime;
41static pthread_mutex_t thread_lock;
42static unsigned int threads_starting;
43static struct stats throughput_stats;
44static pthread_cond_t thread_parent, thread_worker;
45
46struct worker {
47 int tid;
48 u_int32_t *futex;
49 pthread_t thread;
50 unsigned long ops;
51};
52
53static const struct option options[] = {
54 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
55 OPT_UINTEGER('r', "runtime", &nsecs, "Specify runtime (in seconds)"),
56 OPT_UINTEGER('f', "futexes", &nfutexes, "Specify amount of futexes per threads"),
57 OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"),
58 OPT_BOOLEAN( 'S', "shared", &fshared, "Use shared futexes instead of private ones"),
59 OPT_END()
60};
61
62static const char * const bench_futex_hash_usage[] = {
63 "perf bench futex hash <options>",
64 NULL
65};
66
67static void *workerfn(void *arg)
68{
69 int ret;
70 struct worker *w = (struct worker *) arg;
71 unsigned int i;
72 unsigned long ops = w->ops; /* avoid cacheline bouncing */
73
74 pthread_mutex_lock(&thread_lock);
75 threads_starting--;
76 if (!threads_starting)
77 pthread_cond_signal(&thread_parent);
78 pthread_cond_wait(&thread_worker, &thread_lock);
79 pthread_mutex_unlock(&thread_lock);
80
81 do {
82 for (i = 0; i < nfutexes; i++, ops++) {
83 /*
84 * We want the futex calls to fail in order to stress
85 * the hashing of uaddr and not measure other steps,
86 * such as internal waitqueue handling, thus enlarging
87 * the critical region protected by hb->lock.
88 */
89 ret = futex_wait(&w->futex[i], 1234, NULL, futex_flag);
90 if (!silent &&
91 (!ret || errno != EAGAIN || errno != EWOULDBLOCK))
92 warn("Non-expected futex return call");
93 }
94 } while (!done);
95
96 w->ops = ops;
97 return NULL;
98}
99
100static void toggle_done(int sig __maybe_unused,
101 siginfo_t *info __maybe_unused,
102 void *uc __maybe_unused)
103{
104 /* inform all threads that we're done for the day */
105 done = true;
106 gettimeofday(&end, NULL);
107 timersub(&end, &start, &runtime);
108}
109
110static void print_summary(void)
111{
112 unsigned long avg = avg_stats(&throughput_stats);
113 double stddev = stddev_stats(&throughput_stats);
114
115 printf("%sAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n",
116 !silent ? "\n" : "", avg, rel_stddev_stats(stddev, avg),
117 (int) runtime.tv_sec);
118}
119
120int bench_futex_hash(int argc, const char **argv)
121{
122 int ret = 0;
123 cpu_set_t cpuset;
124 struct sigaction act;
125 unsigned int i;
126 pthread_attr_t thread_attr;
127 struct worker *worker = NULL;
128 struct perf_cpu_map *cpu;
129
130 argc = parse_options(argc, argv, options, bench_futex_hash_usage, 0);
131 if (argc) {
132 usage_with_options(bench_futex_hash_usage, options);
133 exit(EXIT_FAILURE);
134 }
135
136 cpu = perf_cpu_map__new(NULL);
137 if (!cpu)
138 goto errmem;
139
140 sigfillset(&act.sa_mask);
141 act.sa_sigaction = toggle_done;
142 sigaction(SIGINT, &act, NULL);
143
144 if (!nthreads) /* default to the number of CPUs */
145 nthreads = cpu->nr;
146
147 worker = calloc(nthreads, sizeof(*worker));
148 if (!worker)
149 goto errmem;
150
151 if (!fshared)
152 futex_flag = FUTEX_PRIVATE_FLAG;
153
154 printf("Run summary [PID %d]: %d threads, each operating on %d [%s] futexes for %d secs.\n\n",
155 getpid(), nthreads, nfutexes, fshared ? "shared":"private", nsecs);
156
157 init_stats(&throughput_stats);
158 pthread_mutex_init(&thread_lock, NULL);
159 pthread_cond_init(&thread_parent, NULL);
160 pthread_cond_init(&thread_worker, NULL);
161
162 threads_starting = nthreads;
163 pthread_attr_init(&thread_attr);
164 gettimeofday(&start, NULL);
165 for (i = 0; i < nthreads; i++) {
166 worker[i].tid = i;
167 worker[i].futex = calloc(nfutexes, sizeof(*worker[i].futex));
168 if (!worker[i].futex)
169 goto errmem;
170
171 CPU_ZERO(&cpuset);
172 CPU_SET(cpu->map[i % cpu->nr], &cpuset);
173
174 ret = pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset);
175 if (ret)
176 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
177
178 ret = pthread_create(&worker[i].thread, &thread_attr, workerfn,
179 (void *)(struct worker *) &worker[i]);
180 if (ret)
181 err(EXIT_FAILURE, "pthread_create");
182
183 }
184 pthread_attr_destroy(&thread_attr);
185
186 pthread_mutex_lock(&thread_lock);
187 while (threads_starting)
188 pthread_cond_wait(&thread_parent, &thread_lock);
189 pthread_cond_broadcast(&thread_worker);
190 pthread_mutex_unlock(&thread_lock);
191
192 sleep(nsecs);
193 toggle_done(0, NULL, NULL);
194
195 for (i = 0; i < nthreads; i++) {
196 ret = pthread_join(worker[i].thread, NULL);
197 if (ret)
198 err(EXIT_FAILURE, "pthread_join");
199 }
200
201 /* cleanup & report results */
202 pthread_cond_destroy(&thread_parent);
203 pthread_cond_destroy(&thread_worker);
204 pthread_mutex_destroy(&thread_lock);
205
206 for (i = 0; i < nthreads; i++) {
207 unsigned long t = worker[i].ops/runtime.tv_sec;
208 update_stats(&throughput_stats, t);
209 if (!silent) {
210 if (nfutexes == 1)
211 printf("[thread %2d] futex: %p [ %ld ops/sec ]\n",
212 worker[i].tid, &worker[i].futex[0], t);
213 else
214 printf("[thread %2d] futexes: %p ... %p [ %ld ops/sec ]\n",
215 worker[i].tid, &worker[i].futex[0],
216 &worker[i].futex[nfutexes-1], t);
217 }
218
219 zfree(&worker[i].futex);
220 }
221
222 print_summary();
223
224 free(worker);
225 free(cpu);
226 return ret;
227errmem:
228 err(EXIT_FAILURE, "calloc");
229}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2013 Davidlohr Bueso <davidlohr@hp.com>
4 *
5 * futex-hash: Stress the hell out of the Linux kernel futex uaddr hashing.
6 *
7 * This program is particularly useful for measuring the kernel's futex hash
8 * table/function implementation. In order for it to make sense, use with as
9 * many threads and futexes as possible.
10 */
11
12/* For the CLR_() macros */
13#include <string.h>
14#include <pthread.h>
15
16#include <errno.h>
17#include <signal.h>
18#include <stdlib.h>
19#include <linux/compiler.h>
20#include <linux/kernel.h>
21#include <sys/time.h>
22
23#include "../util/stat.h"
24#include <subcmd/parse-options.h>
25#include "bench.h"
26#include "futex.h"
27#include "cpumap.h"
28
29#include <err.h>
30
31static unsigned int nthreads = 0;
32static unsigned int nsecs = 10;
33/* amount of futexes per thread */
34static unsigned int nfutexes = 1024;
35static bool fshared = false, done = false, silent = false;
36static int futex_flag = 0;
37
38struct timeval start, end, runtime;
39static pthread_mutex_t thread_lock;
40static unsigned int threads_starting;
41static struct stats throughput_stats;
42static pthread_cond_t thread_parent, thread_worker;
43
44struct worker {
45 int tid;
46 u_int32_t *futex;
47 pthread_t thread;
48 unsigned long ops;
49};
50
51static const struct option options[] = {
52 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
53 OPT_UINTEGER('r', "runtime", &nsecs, "Specify runtime (in seconds)"),
54 OPT_UINTEGER('f', "futexes", &nfutexes, "Specify amount of futexes per threads"),
55 OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"),
56 OPT_BOOLEAN( 'S', "shared", &fshared, "Use shared futexes instead of private ones"),
57 OPT_END()
58};
59
60static const char * const bench_futex_hash_usage[] = {
61 "perf bench futex hash <options>",
62 NULL
63};
64
65static void *workerfn(void *arg)
66{
67 int ret;
68 struct worker *w = (struct worker *) arg;
69 unsigned int i;
70 unsigned long ops = w->ops; /* avoid cacheline bouncing */
71
72 pthread_mutex_lock(&thread_lock);
73 threads_starting--;
74 if (!threads_starting)
75 pthread_cond_signal(&thread_parent);
76 pthread_cond_wait(&thread_worker, &thread_lock);
77 pthread_mutex_unlock(&thread_lock);
78
79 do {
80 for (i = 0; i < nfutexes; i++, ops++) {
81 /*
82 * We want the futex calls to fail in order to stress
83 * the hashing of uaddr and not measure other steps,
84 * such as internal waitqueue handling, thus enlarging
85 * the critical region protected by hb->lock.
86 */
87 ret = futex_wait(&w->futex[i], 1234, NULL, futex_flag);
88 if (!silent &&
89 (!ret || errno != EAGAIN || errno != EWOULDBLOCK))
90 warn("Non-expected futex return call");
91 }
92 } while (!done);
93
94 w->ops = ops;
95 return NULL;
96}
97
98static void toggle_done(int sig __maybe_unused,
99 siginfo_t *info __maybe_unused,
100 void *uc __maybe_unused)
101{
102 /* inform all threads that we're done for the day */
103 done = true;
104 gettimeofday(&end, NULL);
105 timersub(&end, &start, &runtime);
106}
107
108static void print_summary(void)
109{
110 unsigned long avg = avg_stats(&throughput_stats);
111 double stddev = stddev_stats(&throughput_stats);
112
113 printf("%sAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n",
114 !silent ? "\n" : "", avg, rel_stddev_stats(stddev, avg),
115 (int) runtime.tv_sec);
116}
117
118int bench_futex_hash(int argc, const char **argv)
119{
120 int ret = 0;
121 cpu_set_t cpuset;
122 struct sigaction act;
123 unsigned int i;
124 pthread_attr_t thread_attr;
125 struct worker *worker = NULL;
126 struct cpu_map *cpu;
127
128 argc = parse_options(argc, argv, options, bench_futex_hash_usage, 0);
129 if (argc) {
130 usage_with_options(bench_futex_hash_usage, options);
131 exit(EXIT_FAILURE);
132 }
133
134 cpu = cpu_map__new(NULL);
135 if (!cpu)
136 goto errmem;
137
138 sigfillset(&act.sa_mask);
139 act.sa_sigaction = toggle_done;
140 sigaction(SIGINT, &act, NULL);
141
142 if (!nthreads) /* default to the number of CPUs */
143 nthreads = cpu->nr;
144
145 worker = calloc(nthreads, sizeof(*worker));
146 if (!worker)
147 goto errmem;
148
149 if (!fshared)
150 futex_flag = FUTEX_PRIVATE_FLAG;
151
152 printf("Run summary [PID %d]: %d threads, each operating on %d [%s] futexes for %d secs.\n\n",
153 getpid(), nthreads, nfutexes, fshared ? "shared":"private", nsecs);
154
155 init_stats(&throughput_stats);
156 pthread_mutex_init(&thread_lock, NULL);
157 pthread_cond_init(&thread_parent, NULL);
158 pthread_cond_init(&thread_worker, NULL);
159
160 threads_starting = nthreads;
161 pthread_attr_init(&thread_attr);
162 gettimeofday(&start, NULL);
163 for (i = 0; i < nthreads; i++) {
164 worker[i].tid = i;
165 worker[i].futex = calloc(nfutexes, sizeof(*worker[i].futex));
166 if (!worker[i].futex)
167 goto errmem;
168
169 CPU_ZERO(&cpuset);
170 CPU_SET(cpu->map[i % cpu->nr], &cpuset);
171
172 ret = pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset);
173 if (ret)
174 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
175
176 ret = pthread_create(&worker[i].thread, &thread_attr, workerfn,
177 (void *)(struct worker *) &worker[i]);
178 if (ret)
179 err(EXIT_FAILURE, "pthread_create");
180
181 }
182 pthread_attr_destroy(&thread_attr);
183
184 pthread_mutex_lock(&thread_lock);
185 while (threads_starting)
186 pthread_cond_wait(&thread_parent, &thread_lock);
187 pthread_cond_broadcast(&thread_worker);
188 pthread_mutex_unlock(&thread_lock);
189
190 sleep(nsecs);
191 toggle_done(0, NULL, NULL);
192
193 for (i = 0; i < nthreads; i++) {
194 ret = pthread_join(worker[i].thread, NULL);
195 if (ret)
196 err(EXIT_FAILURE, "pthread_join");
197 }
198
199 /* cleanup & report results */
200 pthread_cond_destroy(&thread_parent);
201 pthread_cond_destroy(&thread_worker);
202 pthread_mutex_destroy(&thread_lock);
203
204 for (i = 0; i < nthreads; i++) {
205 unsigned long t = worker[i].ops/runtime.tv_sec;
206 update_stats(&throughput_stats, t);
207 if (!silent) {
208 if (nfutexes == 1)
209 printf("[thread %2d] futex: %p [ %ld ops/sec ]\n",
210 worker[i].tid, &worker[i].futex[0], t);
211 else
212 printf("[thread %2d] futexes: %p ... %p [ %ld ops/sec ]\n",
213 worker[i].tid, &worker[i].futex[0],
214 &worker[i].futex[nfutexes-1], t);
215 }
216
217 free(worker[i].futex);
218 }
219
220 print_summary();
221
222 free(worker);
223 free(cpu);
224 return ret;
225errmem:
226 err(EXIT_FAILURE, "calloc");
227}